Assessing urban flood susceptibility in Seoul, South Korea using machine learning models: effects of urban infrastructure and sampling variability

초록

Urban flooding causes severe damage in densely populated cities. Machine learning (ML) models have recently been adopted to assess urban flood susceptibility. Previous studies have rarely considered changes in ML model performance according to variable composition and repeated sampling. This study evaluated ML model performance for urban flood susceptibility in Seoul, South Korea under different input-data configurations (repeated sampling, and urban infrastructure variables [road, sewer system, and urban detention water system densities]). Two datasets were prepared with and without urban infrastructure using the same topographic and pedological variables. Flood inventory data (2010-2024) were used to define flooded and non-flooded areas, and five repeated samples were generated without replacement. Along with random forest (RF), extreme gradient boost (XGB), multilayer perceptron (MLP), and long short-term memory (LSTM), a transformer-based model, tabular prior-data fitted network (TabPFN), was applied. The major contributing variables were investigated using Shapley additive explanation (SHAP) analysis. The results showed that ML models incorporating urban infrastructure achieved approximately 0.03 higher accuracy and ROC-AUC than those without urban infrastructure. Among ML models, TabPFN showed the highest performance with an average accuracy of 0.81 and ROC-AUC of 0.90. Road density was the most influential variable, followed by elevation. For TabPFN, urban infrastructure inclusion increased the very high and low susceptibility classes by 2.7% and 17.0%, respectively, thereby enhancing spatial agreement with the historical flood data. These findings highlight the importance of urban infrastructure and the value of transformer-based models for urban flood susceptibility assessment.

키워드